[Bugfix][Kernel] Give unique name to BlockSparseFlashAttention (vllm-project#12040)

Signed-off-by: Chen Zhang <[email protected]>

Author: heheda12345

vllm/attention/backends/blocksparse_attn.py

@@ -89,8 +89,7 @@ class BlocksparseFlashAttentionBackend(AttentionBackend):
 
     @staticmethod
     def get_name() -> str:
-        # For attention layer compatibility
-        return "FLASH_ATTN"
+        return "BLOCK_SPARSE_FLASH_ATTN"
 
     @staticmethod
     def get_impl_cls() -> Type["BlocksparseFlashAttentionImpl"]:

vllm/platforms/interface.py

@@ -33,6 +33,7 @@ class _Backend(enum.Enum):
     HPU_ATTN = enum.auto()
     PALLAS = enum.auto()
     IPEX = enum.auto()
+    BLOCK_SPARSE_FLASH_ATTN = enum.auto()
     NO_ATTENTION = enum.auto()